Reinforcement Learning Capability

[rdarnold]

The UAVs should be able to use reinforcement learning to accomplish their mission rather than loading from the simulation matrix. This would be a parallel method to running large-scale experiments. The swarm itself could choose its next configuration, and/or attempt to organize according to various RL algorithms.

[mattosinski]

adding some brainstorming notes:

goal - find the optimal simulation matrix configuration that minimizes rescue time

1. State representation - output of a particular run - rescue time? sector coverage?
2. Action - modifying drone behaviors in simulation_matrix.xlsx
3. Reward - @Reddy2 can you review? One idea could be compare the actual rescue time against a baseline. Or base the reward on if the drone's sensor coverage overlapped a sector that was previously measured?

---

implementation brainstorm:

dronelab controller: create and run an instance of a dronelab

• inputs: custom Simulation_Matrix.xlsx & Config.txt file locations
• output: directory of new dronelab instance

example: controller.go sim="C:\Users\P008702F\Desktop\Simulation_Matrix.xlsx" config="C:\Users\P008702F\Desktop\config.txt"

output: C:\Users\P008702F\dronelab_89df2b2e-fb31-47ba-8dd6-64a33a38e55c